An article tells you how Xiaomi played "artificial intelligence" | CCF-GAIR

Editor's note: On August 12th and 13th, 2016, Lei Fengnet (search "Lei Feng Net" public concern) held an unprecedented "Global Artificial Intelligence and Robotics Summit (CCF-GAIR)" in Shenzhen, from China and the United States and Canada. The most advanced scientific research laboratories, academic authorities, industry leaders, and investment leaders discuss the future trends in the four major areas of artificial intelligence, robotics, drones, and smart driving, and the directions of the development of industry, universities, and research. Lei Fengnet has carefully edited the content of the speech and the video of the conference and will gradually release it in the near future.

Regarding how Xiaomi played "artificial intelligence", the company's co-founder Huang Jiangji made his first comprehensive exposition on the stage of the CCF-GAIR conference of Lei Feng.

His lecture was very engineer-style, and it logically demonstrated the other side of Xiaomi's use of the mass data and machine learning technology produced by the product, landing "artificial intelligence" on various Internet businesses, and promoting the hardware of the Internet of Things. Intelligent.

Unlike those entrepreneurs who can't wait to put the words "AI" on their heads, the entrepreneurs who are very good are very different. The content that Huang Jiangji says is closely integrated with the actual business. It is a typical mature company's approach to the turning point of technology. This is also the most respected idea of ​​Lei Feng.

Huang Jiangji said that Xiaomi’s use of artificial intelligence methods and services (search, recommendation, advertising, cloud albums, etc.) may not be exactly the same as the artificial intelligence AI that everyone imagines, but these businesses benefit from deep learning technologies and The results are immediate.

I think that breakthroughs in artificial intelligence and deep learning over the past few years, and contributions to this industry, are far more than basic techniques such as speech recognition and image processing that everyone understands. In fact, every application and product that needs an algorithm can use it. This is the greatest significance of machine learning to us. It completely changes the methodology of engineers' products.

Huang Jiangji's methodologies refer to the data generated by the use of the product by users, and the processing of machine learning technology makes the products more intelligent, thus forming a positive cycle (see the figure above). He believes that this cycle has become the new normal for product development.

Millet's "Data Factory"

According to Huang Jiangji, Xiaomi’s data mainly comes from mobile phones and other Internet of Things IOT devices (also known as smart devices).

Mobile phones. Milli’s MIUI system has more than 200 million users, of which 8 applications have daily activity exceeding 10 million, including: weather, calendars, app stores, music, photo albums, security centers, browsers, and millet videos; daily activity exceeds 1 million There are 17 applications, in addition to the above-mentioned 8, there are nine, namely: millet life, personality theme, Mi Jia, notes, games, mailboxes, wallets, finance, reading.

Huang Jiangji said:

The activity of the application has two meanings for Xiaomi: First, the real needs of each user are understood through the user's usage feedback; second, because the user is very active in using the mobile phone, Xiaomi can obtain more data. . These data, in turn, help Xiaomi improve the product.

For IoT devices, mobile phones are not included, and Xiaomi has sold more than 37 million smart devices. Three years ago, with the wave of IOT, Xiaomi began to get involved in the hardware used in user life, including televisions, routers, bracelets, and a series of devices from the ecosystem.

Huang Jiangji said:

Xiaomi enters this field and has customer service with two core issues. One is the traditional equipment intelligence. First, it should have networking control capabilities. Xiaomi reduced the cost of the Wi-Fi module from about RMB 60 to about RMB 10; the other is that the networking method is not smart enough. Xiao Mi automatically asks whether to add the device to the home intelligence network after the device is plugged in within his own ecosystem.

“Everyone may have one or two mobile phones, but the number of smart devices in the home may add up to a dozen or more. More than 37 million smart devices are shipped by Xiaomi. In the future, this figure will be much larger than mobile phones.” Huang Jiangji believes that it is because mobile phones and smart devices are actively used by users that they have produced truly high-quality big data.

“The big data of our entire line of products or business is all rolled into one data factory, and the daily data volume is up to 200 TB.” Huang Jiangji said, “The storage cost of data is very high. If you save a lot of data, it cannot be used. If we don't have value, we might as well not have it. We have to process and mine the data so that it can in turn train our algorithms, train a variety of products, and make them truly intelligent."

Millet's "user portrait"

After having a huge amount of data, we realized that we must be human-centered, so it took us two years to establish user portraits. Different companies have different user portraits. In Xiaomi, our definition of user portrait is actually around this user. Whether he is using a mobile phone, watching TV, or using a variety of smart hardware at home, how can Fully transform the needs and habits of this user into the characteristics of the user's portrait.

"Only when you understand his needs better than the users can provide him with more intelligent services." Huang Jiangji gave an example of reading millet. He said that if a user frequently uses millet to read, he is a person who loves reading. Such features are coarse-grained, very representational, and it is difficult to achieve user expectations through such features. Millet needs to know what kinds of books the user likes, what kinds of authors, when they like to read, move, or wait in line for such fragmented times. Only by reaching such a user's image granularity can they provide content recommendation intelligence that exceeds expectation.

Huang Jiangji revealed that Xiaomi's user portrait data already contains about 1,000 user features. Xiaomi also found that with accurate user portrait data, it can also be used to effectively combat cattle. About half a year ago, Xiaomi's engineers tested one hand and judged the user's authenticity through sufficient user characteristics. The Cattle will use one or two "holes" to make the machine think he is a real user. Huang Jiangji said that you can't prevent the cattle from cheating on certain features, but you will find it difficult for the cattle to cheat on 30 or even 100 features at the same time, because the cost of such “treating” is too high. “When we used 1000 features, we found that we could have already lost the cattle.” According to the PPT data, the cattle have fallen from more than 20% to less than 10% (upper right corner in the figure above).

"In addition to getting more and more understanding of users, we have to understand more and more about the content. Xiaomi puts all content resources into the content pool, sorts and extracts features, and finally combines user portraits to provide comprehensive content recommendations. "Huang Jiangji said.

Artificial intelligence is the new turning point

Huang Jiangji said that in the past 6 years, Xiaomi has been fortunate enough to see the turning point and the eruption of several technologies and products from a zero distance: From the outbreak of the mobile Internet, almost all users have used their first smartphone; The IOT broke three years ago; until now artificial intelligence. "I think these outbreaks and inflections are really fortunes for our entire industry."

“Despite experiencing so many technological inflection points, Xiaomi’s goal has never changed. It is still how to make products and functions that users really need more and more useful. We hope to bring these cutting-edge technologies and functions Into the millions of households, not just tall users can use."

We want to work with industry scientists, academic experts, and industrial partners to see if we can place artificial intelligence in our generation and bring AI to every corner of the family. Not at all. I can control the TV without a remote control in the living room. I can tell my smart speaker where to go now and then it orders automatically. This is the first step. The second step, when I didn't say it, was to automatically anticipate, make recommendations, or other smart services. If it can be done, artificial intelligence really does it in everyday life.

"I'm not worried about Singularity. It's not how artificial intelligence rules the world. What I worry about is how, in our generation, industry and the scientific community, we use our technology to bring AI to our lives. Huang Jiangji said finally.

(Appendix: Huang Jiangji's speech record and 35 minutes of video)

Posted on